How to make causal inferences using texts.

Area of Science:

• Social Sciences
• Computational Social Science
• Text Analysis

Background:

• Text-as-data methods offer potential for social science theory testing.
• Causal inferences from text often rely on latent text representations.
• Estimating these representations can introduce identification problems or overfitting.

Purpose of the Study:

• To introduce a rigorous workflow for causal inference with discovered measures.
• To mitigate risks associated with estimating latent text representations.
• To apply the workflow to real-world social science studies.

Main Methods:

• Developed a split-sample workflow for causal inference.
• Applied the workflow to analyze immigration attitudes.
• Utilized the workflow in a study on bureaucratic responsiveness.

Main Results:

• The split-sample workflow enhances the rigor of causal inference from text data.
• Demonstrated the workflow's applicability in social science experiments.
• Provided a method to address underacknowledged risks in text-based causal inference.

Conclusions:

• The proposed split-sample workflow is crucial for reliable text-based causal inference.
• This method improves the validity of social science research using large text collections.
• Researchers can confidently use discovered measures in causal analyses.